OpenAI just shipped Workspace Agents, a feature that lets organizations build custom AI agents inside ChatGPT to run entire workflows without hand-holding. The agents can review leads, summarize support tickets, generate reports, and connect with external tools like ticketing systems and document editors. It's available now in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans.
The enterprise controls are what make this more than a chatbot with tool access. Admins get role-based access control, audit logs, and approval gates that require human sign-off before an agent takes sensitive actions like sending messages or updating records. You can schedule agents for recurring tasks and share a single agent across your workspace so teams follow the same processes. These are the kinds of controls IT departments need before they'll even consider letting AI touch production systems.
But OpenAI's announcement doesn't address cost.
Running agents on capable models like GPT-4o or o1 gets expensive at scale. Tech community discussions suggest companies will likely route most agent workloads through cheaper models like GPT-4o-mini to keep operational expenses down. That means shorter context windows, weaker reasoning, and higher failure rates on complex tasks that plague agents in production. The gap between what these agents promise and what they deliver in production could be real, especially for multi-step workflows that require actual reasoning depth.
This puts OpenAI head to head with Anthropic in the enterprise agent race. AI agents can handle routine workflows just fine. Getting companies to pay for models reliable enough to trust with real work is the actual challenge given the soaring costs of enterprise AI infrastructure.